Scholarlink

<p>Multi-view subspace clustering, which aims to discover latent subspace structures from multiple data sources, has demonstrated superior capability in handling high-dimensional complex datasets. Existing methods primarily construct affinity matrices by exploiting cross-view consistency for clustering analysis, their performance remains constrained by two critical limitations: (1) dependency on initial subspace representation quality due to the absence of refinement and (2) the instability of the view-consistency matrices caused by view-specific noise in complex scenarios. To address these challenges, we propose a novel Multi-view Subspace Clustering (DSE-MVSC) framework with dual structural enhancement mechanisms. Our approach introduces two innovative modules: Structural Separation Information (SSI) Extraction and Structural Reinforcement Information (SRI) Extraction. Specifically, SSI identifies weak inter-sample connections through reconstructed subspace representations, while SRI captures high-order similarity patterns via multi-round <i>k</i>-nearest neighbor (<i>k</i>-NN) constraints with differential sparsity. The final ensemble affinity matrix is synthesized through a voting mechanism that integrates SSI and SRI, effectively mitigating noise interference and model bias. Rigorous evaluations on eight benchmark datasets, comprising text dataset and image dataset, show that DSE-MVSC achieves better performance than eight state-of-the-art methods across multiple clustering metrics, achieving an average improvement of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{18.97\%}\)</EquationSource> </InlineEquation> in accuracy. Ablation studies further validate the complementary roles of SSI and SRI.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-structure enhanced multi-view subspace clustering: a voting-based ensemble approach

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Xiaohuan Lan,
  • Wenbin Gao,
  • Ruibin Ren

摘要

Multi-view subspace clustering, which aims to discover latent subspace structures from multiple data sources, has demonstrated superior capability in handling high-dimensional complex datasets. Existing methods primarily construct affinity matrices by exploiting cross-view consistency for clustering analysis, their performance remains constrained by two critical limitations: (1) dependency on initial subspace representation quality due to the absence of refinement and (2) the instability of the view-consistency matrices caused by view-specific noise in complex scenarios. To address these challenges, we propose a novel Multi-view Subspace Clustering (DSE-MVSC) framework with dual structural enhancement mechanisms. Our approach introduces two innovative modules: Structural Separation Information (SSI) Extraction and Structural Reinforcement Information (SRI) Extraction. Specifically, SSI identifies weak inter-sample connections through reconstructed subspace representations, while SRI captures high-order similarity patterns via multi-round k-nearest neighbor (k-NN) constraints with differential sparsity. The final ensemble affinity matrix is synthesized through a voting mechanism that integrates SSI and SRI, effectively mitigating noise interference and model bias. Rigorous evaluations on eight benchmark datasets, comprising text dataset and image dataset, show that DSE-MVSC achieves better performance than eight state-of-the-art methods across multiple clustering metrics, achieving an average improvement of \(\varvec{18.97\%}\) in accuracy. Ablation studies further validate the complementary roles of SSI and SRI.